xuancaiqisehua/icefall_asr_tal-csasr_conv_emformer_transducer_stateless2

Note: This recipe is trained with the codes from this PR https://github.com/k2-fsa/icefall/pull/874.

Pre-trained conv_emformer_transducer_stateless2 models for the TAL_CSASR dataset with icefall.

The model was trained on the far data of TAL_CSASR with the scripts in icefall based on the latest version k2. You can use the trained model to export it to ncnn and run it with sherpa-ncnn.

Training procedure

• Install k2 : https://k2.readthedocs.io/en/latest/installation/index.html
• Install lhotse : https://lhotse.readthedocs.io/en/latest/getting-started.html#installation
• Clone icefall : https://github.com/k2-fsa/icefall

git clone https://github.com/k2-fsa/icefall
cd icefall

• Preparing data.

cd egs/tal_csasr_conv_emformer/ASR
bash ./prepare.sh

• Training

bash run.sh

• Evaluation results The decoding results (CER%) on TAL_CSASR(dev and test) are listed below:

decoding-method	epoch(iter)	avg	dev	test
fast_beam_search	6	3	11.36	11.37

• Export model to ncnn reference : https://k2-fsa.github.io/icefall/model-export/export-ncnn.html

./conv_emformer_transducer_stateless2/export-for-ncnn.py \
--exp-dir exp_conv_emformer \
--lang_dir data/lang_char \
--epoch 5 \
--iter 8000 \
--avg 3 \
--use-averaged-model 1 \
--num-encoder-layers 12 \
--chunk-length 32 \
--cnn-module-kernel 31 \
--left-context-length 32 \
--right-context-length 8 \
--memory-size 32

• Export torchscript model via pnnx

pnnx ./encoder_jit_trace-pnnx.pt
pnnx ./decoder_jit_trace-pnnx.pt
pnnx ./joiner_jit_trace-pnnx.pt

• Modify the following two lines in your encoder_jit_trace-pnnx.ncnn.param file.

• Then you can use the following code to test the converted models.

model/tokens.txt \
model/encoder_jit_trace-pnnx.ncnn.param \
model/encoder_jit_trace-pnnx.ncnn.bin \
model/decoder_jit_trace-pnnx.ncnn.param \
model/decoder_jit_trace-pnnx.ncnn.bin \
model/joiner_jit_trace-pnnx.ncnn.param \
model/joiner_jit_trace-pnnx.ncnn.bin \
test_wavs/0.wav